Cyclin-dependent kinase 5 (CDK5) has emerged as a principal therapeutic target for Alzheimer's disease. It is highly desirable to develop computational models that can predict the inhibitory effects of a compound towards CDK5 activity. In this study, two machine learning tools (naive Bayesian and recursive partitioning) were used to generate four single classifiers from a large dataset containing 462 CDK5 inhibitors and 1,500 non-inhibitors. Then, two types of consensus models [combined classifier-artificial neural networks (CC-ANNs) and consensus prediction] were applied to combine four single classifiers to obtain superior performance. The results showed that both consensus models outperformed four single classifiers, and (MCC = 0.806) was superior to consensus prediction (MCC = 0.711) for an external test set. To illustrate the practical applications of the CC-ANN model in virtual screening, an in-house dataset containing 29,170 compounds was screened, and 40 compounds were selected for further bioactivity assays. The assay results showed that 13 out of 40 compounds exerted CDK5/p35 inhibitory activities with IC50 values ranging from 9.23 to 229.76 μM. Interestingly, three new scaffolds that had not been previously reported as CDK5 inhibitors were found in this study. These studies prove that our protocol is an effective approach to predict small-molecule CDK5 affinity and identify novel lead compounds.